PMLE Study Guide — How to Pass Google Professional Machine Learning Engineer

Weeks 1–2

Problem Framing and Data Strategy: translating business problems to ML tasks, data collection, bias detection

Tip: Know the difference between ML problem types: supervised (labelled data), unsupervised (unlabelled, find patterns), reinforcement learning (reward-based), and self-supervised (use data itself as labels). Questions describe a business problem and ask which problem type applies and which GCP service supports it.

Weeks 3–5

Data Processing: Vertex AI Datasets, BigQuery ML, Dataflow (Apache Beam), Feature Store

Tip: Vertex AI Feature Store provides a centralised repository for ML features with online serving (low-latency lookups for predictions) and offline serving (batch export for training). Know why Feature Store solves the training-serving skew problem — it ensures the same feature computation is used in both training and prediction.

Weeks 6–9

Model Development: Vertex AI Workbench, AutoML, custom training, hyperparameter tuning

Tip: Vertex AI AutoML vs custom training: AutoML is no-code/low-code, trains on your data using Google-managed architectures, best when you lack ML expertise or have a standard use case. Custom training gives full control over the model architecture and training loop — necessary for research or highly specialised domains.

Weeks 10–14

MLOps: Vertex AI Pipelines, Model Registry, Model Monitoring, explainability

Tip: Vertex AI Model Monitoring detects skew (difference between training data distribution and prediction request distribution) and drift (change in prediction request distribution over time). Know the monitoring job types and what each detects — questions give symptoms of a degraded model and ask which monitor type would catch them.